' ' METHOD AND DEVICE FOR IMPROVING THE ROBUSTNESS AGAINST 'ADVERSARIAL EXAMPLES'

The present invention relates to a method of generating a manipulated data signal (x^adv) for tricking a first machine learning system (60) configured to determine the semantic division (y_cls) of a received one-dimensional or multidimensional data signal (x). The method includes: a step (a) of dete...

Full description

Saved in:
Bibliographic Details
Main Authors MUMMADI CHAITHANYA KUMAR, METZEN JAN HENDRIK, FISCHER VOLKER
Format Patent
LanguageEnglish
Korean
Published 31.10.2018
Subjects
Online AccessGet full text

Cover

Loading…
More Information
Summary:The present invention relates to a method of generating a manipulated data signal (x^adv) for tricking a first machine learning system (60) configured to determine the semantic division (y_cls) of a received one-dimensional or multidimensional data signal (x). The method includes: a step (a) of determining the requested semantic division (y^target) of the manipulated data signal (x^adv); and a step (b) of generating the manipulated data signal (x^adv) in accordance with the estimated semantic division (f_θ;) of the manipulated data signal (x^adv) as well as the received data signal (x) and the determined requested semantic division (y^target). 본 발명은, 수신되는 일차원 또는 다차원 데이터 신호(x)의 시맨틱 분할(y_cls)을 결정하도록 구성되는 제1 머신 러닝 시스템(60)을 기만하기 위한 조작 데이터 신호(x)를 생성하기 위한 방법에 관한 것이며, 상기 방법은 a) 조작 데이터 신호(x)의 요청 시맨틱 분할(y)을 결정하는 단계와; b) 수신된 데이터 신호(x) 및 결정된 요청 시맨틱 분할(y)뿐 아니라 조작 데이터 신호(x)의 추정 시맨틱 분할(f)에 따라서도 조작 데이터 신호(x)를 생성하는 단계를; 포함한다.
Bibliography:Application Number: KR20180024180